Real silicon.
Published datasheet.
Every customer gets the same SKU. No "performance tier" upsell, no mystery instance type names. If you can read a datasheet, you already know what you're getting.
Real silicon.
Not a slide.
Every node is a published datasheet. Same chips, same disks, same network — for everyone.
What's actually in the box.
Compute
Intel Xeon (Cascade Lake) and AMD EPYC (Milan) at launch — pick your CPU class per VM, dial vCPU / RAM / disk independently, all billed per second. Additional CPU families come online with new DCs.
GPU
NVIDIA RTX 4090 via whole-card PCIe passthrough — a dedicated physical GPU, not a time-sliced fraction. The card's vCPU and RAM are reserved alongside it (GPU VMs start at 4 vCPU / 8 GB). Attach one or more per VM; billed per second like everything else. The RTX 4090 is the first card in the catalog — more GPU models roll out as inventory comes online.
Memory
ECC across the fleet. Same memory profile in every host — no surprise downclocking based on which node you land on.
Storage
All-flash NVMe, network-attached so a host failure doesn't take your data with it. Replicated across drives at the storage tier — drive loss is survivable without intervention. On-demand VM snapshots today, billed at 50% of normal NVMe storage. Cron-style schedules + per-disk retention policies ship in v1.x.
Network
50 Gbps per host on a 100 GbE backbone. Native IPv6 /64 per project; public IPv4 is an optional add-on at $2/mo. Bare-metal leases include unlimited bandwidth. VMs get a 10 TB/mo transfer allowance; per-VM and per-network caps when you need them. Traffic between your resources in the same datacenter is free; VM egress to the internet bills against your transfer allowance, and we don't charge once you're past it without saying so.
Get on the metal,
at half price.
Public launch june 22, 2026. Drop your email — first hundred founders get half off everything for the first three months. No credit card. No "schedule a call."
waitlist closes when the first 100 fill